Upload
others
View
3
Download
0
Embed Size (px)
Citation preview
DOCUMENT RESUME
ED 422 359 TM 028 911
AUTHOR Yan, Duanli; Lewis, Charles; Stocking, MarthaTITLE Adaptive Testing without IRT.PUB DATE 1998-04-00NOTE 23p.; Paper presented at the Annual Meeting of the National
Council on Measurement in Education (San Diego, CA, April12-16, 1998).
PUB TYPE Reports Evaluative (142) -- Speeches/Meeting Papers (150)EDRS PRICE MF01/PC01 Plus Postage.DESCRIPTORS *Adaptive Testing; Algorithms; *Computer Assisted Testing;
*Item Response Theory; Models; *Nonparametric Statistics;*Regression (Statistics); Simulation; *Test Construction
ABSTRACTIt is unrealistic to suppose that standard item response
theory (IRT) models will be appropriate for all new and currently consideredcomputer-based tests. In addition to developing new models, researchers willneed to give some attention to the possibility of constructing and analyzingnew tests without the aid of strong models. Computerized adaptive testingcurrently relies heavily on IRT. Alternative, empirically based,nonparametric adaptive testing algorithms exist, but their properties arelittle known. This paper introduces an adaptive testing algorithm thatbalances maximum differentiation among test takers with stable estimation ateach stage of testing, and compares this algorithm with a traditional oneusing IRT and maximum information. The adaptive testing algorithm introducedis based on the classification and regression tree approach described in L.Breiman, J. Friedman, R. Olshen, and C. Stone (1984) and J. Chambers and T.Hastie (1992). Simulation results from the regression tree approach werecompared with simulation results from three parameter logistic model IRT.Simulation results show that the nonparametric tree-based approach toadaptive testing may be superior to conventional IRT-based adaptive testingin cases where the IRT assumptions are not satisfied. It clearly outperformedone-dimensional IRT when the pool was strongly two-dimensional. A technicalappendix describes the algorithm. (Contains three figures and sixreferences.) (SLD)
********************************************************************************* Reproductions supplied by EDRS are the best that can be made *
* from the original document. *
********************************************************************************
U.S. DEPARTMENT OF EDUCATIONOffice of Educational Research and Improvement
EDU ATIONAL RESOURCES INFORMATIONCENTER (ERIC)
This document has been reproduced asreceived from the person or organizationoriginating it.
0 Minor changes have been made toimprove reproduction quality.
Points of view or opinions stated in thisdocument do not necessarily representofficial OERI position or policy.
CO
2
Adaptive Testing without IRT
Duanli Yan, Charles Lewis and Martha StockingEducational Testing Service
Princeton, NJ
A
PERMISSION TO REPRODUCE ANDDISSEMINATE THIS MATERIAL
HAS BEEN GRANTED BY
cui
TO THE EDUCATIONAL RESOURCESINFORMATION CENTER (ERIC)
It is unrealistic to suppose that standard item response (MT) models will be
..appropriate for all the new and currently considered computer-based tests . In addition to
developing new models, we also need to give some attention to the possibility of
constructing and analyzing new tests without the aid of strong models. Computerized
adaptive testing currently relies heavily on item response theory. Alternative, empirically
based, non-parametric adaptive testing algorithms exist, but their properties are little
known.
Introduction
Wainer et al. (1991) and Wainer et al. (1992) introduced a testlet-based algebra
exam and compared a hierarchically constructed (adaptive) 4-item testlet with a linear
(fixed format) testlet under various conditions. They compared an adaptive test using an
optimal 4-item tree and a best fixed 4-item test (both defined in terms of maximum
differentiation) through cross-validation, and found that overall the adaptive testlet
dominates the best fixed testlet, but the superiority (at a considerable cost for adaptive
testlet over fixed testlet) is modest. They also found that the adaptive test outperforms the
fixed test for the groups with extreme observed scores. They concluded that, for
circumstances similar to their cases, "a fixed format testlet that uses the best items in the
pool can do almost as well as the optimal adaptive testlet of equal length from that same
pool."
Paper presented at the annual meeting of the National Council on Measurement inEducation, San Diego, CA, April 1998.
2
CAT without IRT
Page 2
Schnipke and Green (1995) compared an item selection algorithm based on
maximum differentiation among test takers with one using item response theory and based
on maximum information. Overall, adaptive tests based on maximum information provided
the most information over the widest range of ability values and, in general, differentiated
among test takers slightly better than the other tests. Although the maximum
differentiation technique may be adequate in some circumstances, adaptive tests based on
maximum information were clearly superior in their study.
This paper introduces an adaptive testing algorithm that balances maximum
differentiation among test takers with stable estimation at each stage of testing, and
compares this algorithm with a traditional one using item response theory and maximum
information.
Method
In this paper, we consider adaptive testing as a prediction system. Specifically, we
use adaptive testing to predict the observed scores that test takers would have received if
they had taken every item in a reference test or a pool. (We restrict our attention to binary
items, scored correct or incorrect.) This is a non-parametric approach in the sense that we
do not introduce latent traits or true scores. We are only considering the observed
number-correct scores test takers would have received if they had taken every item we
could have given. In other words, our criterion is the total observed score for a pool or
reference test. The adaptive testing algorithm we introduce in this paper is based on the
classification and regression tree approach described in Breiman, Friedman, Olshen and
Stone (1984) and in Chambers and Hastie (1992).
In order to construct an adaptive test as a prediction system, we need to have a
calibration sample. Specifically, we need a sample of test takers who take every item in
3
CAT without IRT
Page 3
the pool that will be used for adaptive testing. (For operational use, incomplete
calibration designs would obviously be necessary.) We can then compute the criterion
(total observed score) for these test takers. This is analogous to the calibration sample
one needs when using IRT to do adaptive testing. However, the purpose of the IRT
calibration sample is to calibrate items. Our purpose is not to calibrate items individually,
but to generate a regression tree.
Figure 1 is an example of such a regression tree. The vertical axis represents the
stage of testing and the horizontal axis identifies the prediction of the total score at each
stage. In this example, there are 9 stages (i.e., each test taker would be administered 9
items). The nodes of the tree are plotted as octagons with item numbers inside. The
branches represent the paths test takers could follow in the test, taking the right branch
after answering the item in the octagon correctly and the left branch after answering
incorrectly. At the end, the locations of the terminal nodes, or leaf nodes, give the final
predictions of test takers' total scores.
Once the regression tree has been constructed (and validated -- see below), it may
be used to administer an adaptive test. Thus, based on Figure 1, all test takers would first
be administered item 31. Test takers answering correctly would receive item 27. Those
answering incorrectly would get item 28. Test takers continue through the tree to the
terminal nodes and receive the corresponding final predicted total score as their score on
the test. For instance, test takers who receive item 5 as the last item, and answer it
correctly, would have a predicted total score of 32.5.
Returning to the construction of the tree, suppose we have a calibration sample of
test takers who answered every item in a pool of items. The total number of correct
responses for each test taker is the criterion we will use. Our regression tree begins with
the item (in Figure 1, item 31) that best predicts the observed score in a least squares
4
CAT without IRT
Page 4
sense for these test takers. It splits the calibration sample into two subsamples: those test
takers who answered the item incorrectly and those who answered correctly. They are
represented as nodes in Figure 1: the left node with number 28 inside and right node with
number 27 inside. These two subsamples have maximum differentiation between them
(i.e., maximum sum of squares between subsamples). The horizontal locations of the
nodes are the mean total scores for these two sub-samples.
The construction of the tree continues by finding the best predicting item for those
test takers responding correctly to the first item (in Figure 1, item 27), as well as the best
item for those with an incorrect response to the first item (in Figure 1, item 28). At each
stage, the total calibration sample is split into subsamples and an optimal item is chosen for
each subsample. At each stage, it is also the case that subsamples with similar average
criterion scores are combined as the tree progresses to keep the total number of such
subsamples within reasonable limits. In Figure 1, the nodes with test takers who correctly
answered item 28 and test takers who incorrectly answered item 27 are combined, and the
combined subsamples are administered item 16. At the end of the process, the adaptive
test score given to each test taker is the average criterion score for the final subsample to
which the test taker has been classified (in Figure 1, the combined leaf nodes).
A portion of the prediction for the calibration sample capitalizes on chance. To
evaluate the procedure, we construct the regression tree in a calibration sample and apply
the predictions from the calibration sample to compare with the observed scores in an
application sample. This application sample has the same structure as the calibration
sample. In other words, every test taker answers every item, so a criterion observed score
can be computed. The precision of estimation using the regression tree as an adaptive test
may be measured using the mean of the squared discrepancies (or residuals) between
predicted and observed scores in the application sample. For purposes of interpretation,
5
CAT without IRT
Page 5
this quantity may be compared to the variance of the observed scores in the application
sample. In particular, we will consider the proportion of variance accounted for by the
tree-based predictions.
Results
We compared simulation results from the regression tree approach with
simulations using 3PL IRT and maximum information for the same situations. For our
first set of simulations, we constructed our calibration sample using the 3PL IRT model to
generate item responses for a sample of 500 simulated test takers to 494 items in an
actual item pool for an operational computer adaptive test assessing quantitative
reasoning. (Specifically, we used the 3PL IRT model with item parameters set equal to the
estimates from the operational pool.)
We constructed a regression tree as described in the method section for a 19-item
adaptive test for this calibration sample. The mean of the squared residuals between
predictions and total observed scores for the calibration sample is 11.7. This quantity may
be compared to the variance of the total observed scores for the sample, or 6,586.0. Thus
99.8% of the total observed score variance is accounted for in the calibration sample using
the predictions from the regression tree.
Next, we used the regression tree predictions based on the calibration sample to
compare with the total (MT-based) true scores (rather than total observed scores) in an
application sample of size 10,000, constructed in the same way as the calibration sample.
The mean squared residual in the application sample, based on the calibration predictions
at the end of the 19-item test is 1223.7 (with original true score variance of 6,457.5),
which means the predictions account for 81.0% of the total true score variance. From this
result, we see that there was substantial capitalization on chance in the calibration sample.
6
CAT without IRT
Page 6
From the same calibration sample we used to construct the regression tree, we also
obtained 3PL item parameter estimates using PARSCALE (Muraki & Bock, 1993). We
then carried out an IRT-based (maximum information) adaptive testing simulation on the
application sample using these estimated item parameters. As estimates of total true
score, we used maximum likelihood estimates of the latent trait, transformed using the test
characteristic curve for the entire pool. The mean squared residual between these
'estimates and the total true scores in the application sample is 514.5. Comparing this with
the total true score variance for the application sample, we see that the IRT-based
estimates account for 92.0% of that variance, substantially more than when using the tree-
based predictions. Figure 2 provides a more detailed comparison of the regression tree
and IRT-based CATs as a function of test length.
For our second set of simulations, we considered what would happen if the 3PL
IRT assumptions were violated. We used the same pool, but split it into two equal parts
such that half of the items were considered to measure one latent trait and the other half
measured a second, uncorrelated, latent trait. The parameters for all items were left
unchanged. A calibration sample consisting the responses to all items in the pool for a
sample of 500 simulated test takers was generated, based on the two dimensional latent
trait model just described. Specifically, for each simulated test taker, two latent trait
values were sampled. One of these was used as the basis for response generation for items
in the first half of the pool, while the other was used to generate responses to items in the
second half of the pool. We used the resulting data as our calibration sample for both the
tree-based approach and the one-dimensional IRT model, as before. (Note that the 3PL
model was fit simultaneously to all items in the pool, ignoring which half they were in.)
We also generated responses to all items in the pool for a new sample of 10,000
simulated test takers in the manner just described for use as our application sample. We
7
CAT without IRT
Page 7
used the tree-based predictions from the calibration sample in the application sample for
the regression tree approach. We also carried out a (one-dimensional) IRT adaptive
testing simulation for this application sample, using the item parameters obtained from the
calibration sample. The two-dimensional, IRT-based total true scores for the pool served
as the evaluation criterion for both procedures. Specifically, we compared the mean
squared residuals obtained for the two methods.
Fitting a regression tree to the data for the calibration sample produced a mean
squared residual of 20.8, compared with a total observed score variance of 3,025.9. In
other words, 99.3% of the observed score variance can be accounted for in this calibration
sample using the predictions from the regression tree. (Note that the total observed score
variance in this sample is much smaller than that obtained in the calibration sample based
on the one dimensional IRT model: 3,025.9 compared to 6,586.0. This is a result of the
fact that the between-set item correlations in our second design are all zero.)
In the application sample, using the tree-based predictions from the calibration
sample to predict the total true scores gave a mean squared residual of 1,187.0. The total
true score variance in the application sample is 3,180.0, so 62.7% of this variance is
accounted for by the tree-based predictions. (Here we see an even more substantial
chance capitalization than in the one-dimensional case.) True score estimates based on a
3PL IRT CAT produced a mean squared residual of 1,960.0 in the application sample, so
that only 38.4% of the total true score variance is accounted for by these estimates.
Figure 3 provides a more detailed comparison of the regression tree and IRT-based CATs
as a function of test length.
Discussion
8
CAT without MT
Page 8
For our one-dimensional example, Figure 2 shows that, once the adaptive test is
long enough, the IRTTbased CAT produces consistently better estimates of true scores
than does the tree-based approach, However, it is worth noting that, in the early stages of
testing, the maximum,likelihood estimates from the IRT-based CAT are very poor
compared to those from the regression tree. This suggests a possible hybrid algorithm,
using a regression tree to select the first few items on an adaptive test, and then switching
to a maximum information rRT-based algorithm. This leaves open the question of how
best to make the transition from regression tree to maximum likelihood estimates.
In our two-dimensional example, we see from Figure 3 that the regression tree
clearly provides better prediction than the IRT-based CAT at all test lengths. However, it
should be noted that our example is based on an extreme version of a two-dimensional
model in which every item measures either one or the other dimension (but not both), and
the two uncorrelated dimensions are taken to be equally important.
One of the limitations of the tree-based approach described in this paper is that
there is no control of item exposure rates. (For instance, our algorithm now has everyone
take the same first item.) Another limitation is that no attempt is made to control the
content of the adaptive tests. A third limitation is that all test takers in the calibration and
application samples were assumed to have answered all items in the pool. (All these
limitations also apply to the IRT-based algorithm we used for comparison purposes in this
study. It should be noted, however, that operational MT CATs have none of these
limitations.) Future research will address these and related issues.
9
CAT without IRT
Page 9
Conclusions
We have developed a non-parametric, tree-based approach to adaptive testing and
shown that it may be superior to conventional, IRT-based adaptive testing in cases where
the TRT assumptions are not satisfied. In particular, we showed that the tree-based
approach clearly out-performed (one-dimensional) IRT when the pool was strongly two-
dimensional.
Educational Importance
With new types of computer-based tests being considered, it is important to have
new psychometric tools available to be used with these tests. The non-parametric, tree-
based approach to adaptive testing described here is one such tool. We expect it to be
particularly useful when items test more complex domains than is currently the case with
IRT-based adaptive testing.
1 0
CAT without IRT
Page 10
Technical Appendix - Description of algorithm
Our regression trees are constructed as follows. For each node, we select an
unused item that gives the maximum differentiation (in a least squares sense) on the
criterion score for splitting the current node into two nodes. For each stage, we compare
all the nodes at that stage by computing the pairwise t-statistics and effect size measures,
using the criterion score. If, for some pair of nodes, the absolute value of the t-statistic is
less then some pre-set critical value, or the absolute value of the effect size measure is less
than some pre-set critical value, then we combine the two nodes. (If there is more than
one pair of nodes that meets either of these criteria, we start by combining the pair with
the smallest t-statistic (or smallest effect size if no t-statistic is less than the critical value).
We then recompute all t-statistics and effect sizes for this new node with the others and
repeat the process until all pairs of nodes are distinct in terms of their t-statistics and effect
sizes.
We continue constructing the regression tree stage by stage in this manner until a
specified fixed test length is reached. At the final stage, each test taker in a sample is
classified to one of the leaf nodes based on matching their response pattern with the
regression tree structure. The prediction of that individual's criterion score is the average
score of the leaf node in the regression tree to which the individual is classified.
Exhibit A reproduces an edited version of a portion of the output from the
program we use to construct regression trees. Specifically, it is the output describing the
construction of the tree illustrated in Figure 1. The information given in line 007 describes
the complete calibration sample (node 0) as having 250 (simulated) test takers, a mean
criterion score of 34.7360 and a sum of squared deviations of individual scores around this
mean (Deviance) of 28608.5760. Line 012 repeats some of this information, and notes
1 1
CAT without IRT
Page 11
that item 31 has been selected as the first item in the tree. The remaining output in lines
015 - 023 will be of more interest at later stages.
Lines 027 and 028 describe nodes 1 and 2, defined as those test takers who answer
item 31 incorrectly or correctly, respectively. Specifically, there are 71 of the former and
179 of the latter, with mean criterion scores of 25.3521 and 38.4581, respectively. Lines
029 and 030 give the t-statistic and effect size measure used to compare nodes 1 and 2.
Both values exceed their respective criteria, so no combining of nodes occurs at this stage.
Lines 035 and 036 indicate that items 28 and 27 have been chosen for nodes 1 and 2,
respectively. Line 039 gives the total within-node sum of squares at stage 1 as
19876.6329. Note that this is the sum of the sums of squares for each of the two nodes at
this stage. Line 041 gives the proportion of variance accounted for at this stage, obtained
by subtracting the ratio of the deviance at this stage to the deviance at stage 0 from unity.
Lines 047 and 048 report the standard deviations for nodes 1 and 2.
Lines 052 - 055 describe the four nodes at stage 2 defined by incorrect and correct
answers to items 28 and 27. Lines 056 - 062 give all pairwise comparisons for the nodes
at this stage, as well as the comparison with the smallest t-statistic (obtained for nodes 4
and 5). Since this value (-1.2691) is less in absolute value than our critical value of 2.0,
these two nodes are combined. The new nodes are described in lines 068 - 070, and the
comparisons are given in lines 073-076. No further combination is indicated, so items are
chosen for each of these nodes (2, 16, and 33, respectively), and the final description is
given in lines 078 - 095. The actual output continues in this fashion until the specified
number of stages (test length) has been reached.
12
CAT without IRT
Page 12
References
Breiman, L., Friedman, J. H., Olshen, R. A., & Stone, C. J. (1984). Classification and
regression trees. Pacific Grove, CA: Wadsworth & Brooks/Cole Advanced
Books & Software.
Chambers, J. M., & Hastie, T. J. (1992). Statistical models.. Pacific Grove, CA:
Wadsworth & Brooks/Cole Advanced Books & Software.
Muraki, E., & Bock, D. (1993). PARSCALE: IRT based test scoring and item analysis
for graded open-ended exercises and perfonnance tasks. [Computer program].
Chicago, IL: Scientific Software, Inc.
Schnipke, D., & Green, B. (1995). A comparison of item selection routines in linear and
adaptive tests. Journal of Educational Measurement, 32, 227-242.
Wainer, H., Lewis, C., Kaplan, B., & Braswell, J. (1991). Building algebra testlets: A
comparison of hierarchical and linear structures. Journal of Educational
Measurement, 28, 311-324.
Wainer, H., Kaplan, B., & Lewis, C. (1992). A comparison of the performance of
simulated hierarchical and linear testlets. Journal of Educational Measurement,
29, 243-251.
13
Exh
ibit
A. S
ampl
e O
utpu
t fro
m P
rogr
am to
Con
stru
ct R
egre
ssio
n T
rees
001
002
003
004
005
006
007
008
009
010
011
012
013
014
015
016
017
018
019
020
021
022
023
024
025
026
027
028
029
030
031
032
033
034
035
1 4
036
037
038
ADAPTIVE TESTING WITHOUT IRT - Calibration Sample Tree Table
NodeInfo:istage=
Stage - 0
0 inode=
0 Nsubj= 250 yval= 34.7360 Dev=28608.5760
Stage Node
Item
Parents
Deviance
Mean
00
31
-9,-9,-9,-9,-9
250
28608.5760
34.7360
999.0000
999.0000
Total Deviance at Stage
0:
28608.5760
Proportion of Variance Accounted For:
0.0000
Standard Deviation at root:
10.6974
Standard Deviation for Each Node at This Stage:
Node =
0SD =
10.6974
Stage =
1
NodeInfo:istage=
1 inode=
1 Nsubj=
71 yval=
25.3521 Dev=3716.1972
NodeInfo:istage=
1 inode=
2 Nsubj=
179 yval=
38.4581 Dev=16160.4358
In Combine:inode,jnode,t,e =
12
-11.6993
-1.5452
Inode,Jnode,tmin,emin =
12
-11.6993
-1.5452
Stage Node
Item
Parents
Deviance
Mean
11
28
0,-9,-9,-9,-9
71
3716.1972
25.3521
-11.6993
1.5452
12
27
0,-9,-9,-91-9
179
16160.4358
38.4581
999.0000
999.0000
15
1 G
Exh
ibit
A. S
ampl
e O
utpu
t fro
m P
rogr
am to
Con
stru
ct R
egre
ssio
n T
rees
(C
ontin
ued)
039
040
041
042
043
044
045
046
047
048
049
050
051
052
053
054
055
056
057
058
059
060
061
062
063
064
065
066
067
068
069
070
071
072
073
074
075
076
077
Total Deviance at Stage
1:
19876.6329
Proportion of Variance Accounted For:
0.3052
Standard Deviation at root:
10.6974
Standard Deviation for Each Node at This Stage:
Node =
1SD =
Node =
2SD =
7.2347
9.5017
Stage = 2
NodeInfo:istage=
2 inode=
NodeInfo:istage=
2 inode=
NodeInfo:istage=
2 inode=
NodeInfo:istage=
2 inode=
In Combine:inode,jnode,t,e =
In Combine:inode,jnode,t,e =
In Combine:inode,jnode,t,e =
In Combine:inode,jnode,t,e =
In Combine:inode,jnode,t,e =
In Combine:inode,jnode,t,e =
Inode,Jnode,tmin,emin =
3 Nsubj=
44 yval= 22.8409 Dev=1069.8864
4 Nsubj=
27 yval= 29.4444 Dev=1916.6667
5 Nsubj=
70 yval= 31.7857 Dev=3251.7857
6 Nsubj= 109 yval= 42.7431 Dev=7790.8073
34
-3.6375
-0.9405
35
-8.0368
-1.4907
36
-17.9650
-2.8575
45
-1.2691
-0.3012
46
-7.2206
-1.5573
56
-9.4833
-1.4190
45
-1.2691
-0.3012
Estimations during combination:
Stage Node
Item
Parents
23
24
25
1,-9,-9,-9,-9
1, 2,-9,-9,-9
2,-9,-9,-9,-9
In Combine:inode,jnode,t,e =
In Combine:inode,jnode,t,e =
In Combine:inode,jnode,t,e =
Inode,Jnode,tmin,emin =
Deviance
Mean
44
1069.8864
22.8409
97
5275.2577
31.1340
109
7790.8073
42.7431
34
-7.7947
-1.3126
35
-17.9650
-2.8575
45
-10.4748
-1.4563
34
-7.7947
-1.3126
999.0000
999.0000
-1.2691
0.3012
999.0000
999.0000
17
15
Exh
ibit
A. S
ampl
e O
utpu
t fro
m P
rogr
am to
Con
stru
ct R
egre
ssio
n T
rees
(C
ontin
ued)
078
Stage Node
Item
Parents
NDeviance
Mean
079
080
23
21,-9,-9,-9,-9
44
1069.8864
22.8409
-7.7947
1.3126
081
24
16
1, 2,-9,-9,-9
97
5275.2577
31.1340
999.0000
999.0000
082
25
33
2,-9,-9,-9,-9
109
7790.8073
42.7431
999.0000
999.0000
083
084
085
Total Deviance at Stage
2:
14135.9514
086
087
Proportion of Variance Accounted For:
0.5059
088
089
Standard Deviation at root:
10.6974
090
091
Standard Deviation for Each Node at This Stage:
092
093
Node =
3SD =
4.9311
094
Node =
4SD =
7.3746
095
Node =
5SD =
8.4543
19
1 2 3 4 5 6 7 8 9
Fig
ure
1.R
egre
ssio
n T
ree
Str
uctu
re
1416
1820
2224
-26
2830
3234
3638
4042
44
4648
5052
5456
a 0
Pre
dict
ed S
core
21
Figure 2
Comparison of Tree-based and IRT CATs in One-dimensional Application Sample(referring to true scores)
I = IRT, T = Tree
0 1 2 3 4 5 6 7 8 9 10 12 14 16 18 20
Test Length2 2
Figure 3
Comparison of Tree-based and IRT CATs in Two-dimensional Application Sample(referring to true scores)
I = IRT, T = Tree
0 1 2 3 4 5 6 7 8 9 10 12 14 16 18 20
Test Length2 3
U.S. DepartmentofEducationOffice of Educational Research and Improvement (OERI)
National Library of Education (NLE)Educational Resources Information Center (ERIC)
REPRODUCTION RELEASE(specific:pocumen)
I. DOCUMENT IDENTIFICATION:
ERITM028911
Title:Adaptive Testing without IRT
Author(s): Yan, Duanli, Lewis, Charles and Stocking, Martha
Corporate Source: Publication Date:
II. REPRODUCTION RELEASE:In order to disseminate as widely as possible timely and significant materials of interest to the educational community, documents announced in the
monthly abstract journal of the ERIC system, Resources in Education (RIE), are usually made available to users in microfiche, reproduced paper copy,and electronic media, and sold through the ERIC Document Reproduction Service (EDRS). Credit is given to the source of each document, and, ifreproduction release is granted, one of the following notices is affixed to the document.
If permission is granted to reproduce and disseminate the identified document, please CHECK ONE of the following three options and sign at the bottomof the page.
The sample sticker shown below will beaffixed to all Level 1 documents
1
PERMISSION TO REPRODUCE ANDDISSEMINATE THIS MATERIAL HAS
BEEN GRANTED BY
TO THE EDUCATIONAL RESOURCESINFORMATION CENTER (ERIC)
Level 1
Check here for Level 1 release, permitting reproductionand dissemination In microfiche or other ERIC archival
media (e.g., electronic) and paper copy.
Signhere,--)please
The sample sticker shown below will beaffixed to all Level 2A documents
PERMISSION TO REPRODUCE ANDDISSEMINATE THIS MATERIAL IN
MICROFICHE, AND IN ELECTRONIC MEDIAFOR ERIC COLLECTION SUBSCRIBERS ONLY,
HAS BEEN GRANTED BY
2A
\etc\C
TO THE EDUCATIONAL RESOURCESINFORMATION CENTER (ERIC)
Level 2A
Check here for Level 2A release, permitting reproductionand dissemination In microfiche and in electronic media
for ERIC archival collection subscribers only
The sample sticker shown'below will beaffixed to all Level 2B documents
PERMISSION TO REPRODUCE ANDDISSEMINATE THIS MATERIAL IN
MICROFICHE ONLY HAS BEEN GRANTED BY
2B
cpTO THE EDUCATIONAL RESOURCES
INFORMATION CENTER (ERIC)
Level 28
Check here for Level 2B release, permittingreproduction and dissemination In microfiche only
Documents will be processed as indicated provided reproduction quality permits.If permission to reproduce Is granted, but no box is checked, documents will be processed at Level 1.
I hereby grant to the Educational Resources Information Center (ERIC) nonexclusive permission to reproduce and disseminate this documentas indicated above. Reproductidn from the ERIC microfiche or electronic media by persons other than ERIC employees and its systemcontractors requires permission from the copyright holder. Exception is made for non-profit reproduction by libraries and other service agenciesto satisfy information needs of educators in response to discrete inquiries.
Signature:
7lt491./t()
7j,n1Ad:Areak-c-- eiJ)fr C.G1-041, IV- o
Printed Name/Position/Title:
13er-ne:173 -_59zp4' c:Aex>4. oft-5 c4?
Dale: V7/99iFsmfT. o 12..c.(over)
III. DOCUMENT AVAILABILITY INFORMATION (FROM NON-ERIC SOURCE):
If permission to reproduce is not granted to ERIC, or, if you wish ERIC to cite the availability of the document from another source, pleaseprovide the following information regarding the availability of the document. (ERIC will not announce a document unless it is publiclyavailable, and a dependable source can be specified. Contributors should also be aware that ERIC selection criteria are significantly morestringent for documents that cannot be made available through EDRS.)
Publisher/Distributor:
Address:
Price:
IV. REFERRAL OF ERIC TO COPYRIGHT/REPRODUCTION RIGHTS HOLDER:
If the right to grant this reproduction i.elease is held by someone other than the addressee, please provide the appropriate name andaddress:
Name:
Address:
V. WHERE TO SEND THIS FORM:
Send this form to the following ERIC Clearinghouse:THE UNIVERSITY OF MARYLAND
ERIC CLEAMNGHOUSE ON ASSESSMENT AND EVALUATION ,1129 SHRIVER LAB, CAMPUS DRIVE
COLLEGE PARK, MD 20742-5701Attn: Acquisitions
However, if solicited by the ERIC Facility, or if making an unsolicited contribution to ERIC, return this form (and the document beingcontributed) to:
ERIC Processing and Reference Facility1100 West Street, rd Floor
Laurel, Maryland 20707-3598
Telephone: 301-497-4080Toll Free: 800-799-3742
FAX: 301-953-0263e-mail: [email protected]
WWW: http://ericfac.piccard.csc.com
EFF-088 (Rev. 9/97)PREVIOUS VERSIONS OF THIS FORM ARE OBSOLETE.